Build for Zen5 fails

[AaronBeier]

Trying to build with

cmake -B build -DBLIS_CONFIG_FAMILY=zen5 -DENABLE_CBLAS=ON -DENABLE_THREADING=openmp -DENABLE_ADDON="aocl_gemm" -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release

on and for a Ryzen 9 9900X fails with the following error:

In file included from /home/aaron/code/blis/kernels/zen4/3/sup/bli_dgemmsup_rv_zen4_asm_24x8m.c:36:
/home/aaron/code/blis/kernels/zen4/3/sup/bli_dgemmsup_rv_zen4_asm_24x8m.c: In function ‘bli_dgemmsup_rv_zen4_asm_24x8m_new’:
/home/aaron/code/blis/frame/include/bli_x86_asm_macros.h:102:21: error: ‘asm’ operand has impossible constraints or there are not enough registers
  102 | #define BEGIN_ASM() __asm__ volatile (
      |                     ^~~~~~~
/home/aaron/code/blis/frame/include/bli_x86_asm_macros.h:154:21: note: in expansion of macro ‘BEGIN_ASM’
  154 | #define begin_asm() BEGIN_ASM()
      |                     ^~~~~~~~~
/home/aaron/code/blis/kernels/zen4/3/sup/bli_dgemmsup_rv_zen4_asm_24x8m.c:48382:9: note: in expansion of macro ‘begin_asm’
48382 |         begin_asm()
      |         ^~~~~~~~~
compilation terminated due to -Wfatal-errors.
make[2]: *** [kernels/CMakeFiles/zen4_KERNELS.dir/build.make:625: kernels/CMakeFiles/zen4_KERNELS.dir/zen4/3/sup/bli_dgemmsup_rv_zen4_asm_24x8m.c.o] Error 1
make[1]: *** [CMakeFiles/Makefile2:1469: kernels/CMakeFiles/zen4_KERNELS.dir/all] Error 2
make: *** [Makefile:136: all] Error 2

Full build log: https://gist.github.com/AaronBeier/42c259ce442204cfcd1ff70102dedb0d
Let me know if you need any additional info!

[edwsmyth]

Hi Aaron, this is a problem with gcc 15 and we are working on a solution. In the meantime, is it possible for you to switch back to gcc 14.3 or an earlier release?

[AaronBeier]

Im happy to wait for the fix rather than downgrading gcc, no rush on my end. If youd find it helpful for debugging/testing purposes id be willing to switch back to gcc 14 temporarily. otherwise ill just keep an eye on this issue, thanks you for your work!

[TheTrustedComputer]

I encountered the same error when building for a Zen 3 target on GCC 15.1, 5.1 release. Switching to Clang 20 worked.

[77/663] Building C object kernels/CMakeFiles/zen_KERNELS.dir/zen/3/sup/bli_gemmsup_rv_zen_asm_s6x16m.c.o
FAILED: kernels/CMakeFiles/zen_KERNELS.dir/zen/3/sup/bli_gemmsup_rv_zen_asm_s6x16m.c.o 
/usr/sbin/cc -DBLIS_IS_BUILDING_LIBRARY -DBLIS_VERSION_STRING="\"AOCL-BLAS 5.1.0 Build 20250727\"" -D_POSIX_C_SOURCE=200112L -I/root/rocm-build/blis/include/zen3 -I/root/blis/frame/3 -I/root/blis/frame/1m -I/root/blis/frame/1f -I/root/blis/frame/1 -I/root/blis/frame/include -march=native -O3 -fno-stack-protector -funroll-loops -fno-math-errno -fno-trapping-math -ffp-contract=on -ffunction-sections -fdata-sections -Wl,--gc-sections -s -w -std=c99 -fPIC -O3 -fomit-frame-pointer -mavx2 -mfpmath=sse -mfma -march=znver3 -Wall -Wno-unused-function -Wfatal-errors -fvisibility=hidden -fmacro-prefix-map=/root/blis/= -MD -MT kernels/CMakeFiles/zen_KERNELS.dir/zen/3/sup/bli_gemmsup_rv_zen_asm_s6x16m.c.o -MF kernels/CMakeFiles/zen_KERNELS.dir/zen/3/sup/bli_gemmsup_rv_zen_asm_s6x16m.c.o.d -o kernels/CMakeFiles/zen_KERNELS.dir/zen/3/sup/bli_gemmsup_rv_zen_asm_s6x16m.c.o -c /root/blis/kernels/zen/3/sup/bli_gemmsup_rv_zen_asm_s6x16m.c
In file included from /root/blis/kernels/zen/3/sup/bli_gemmsup_rv_zen_asm_s6x16m.c:39:
/root/blis/kernels/zen/3/sup/bli_gemmsup_rv_zen_asm_s6x16m.c: In function ‘bli_sgemmsup_rv_zen_asm_6x8m’:
/root/blis/frame/include/bli_x86_asm_macros.h:102:21: error: ‘asm’ operand has impossible constraints or there are not enough registers
  102 | #define BEGIN_ASM() __asm__ volatile (
      |                     ^~~~~~~
/root/blis/frame/include/bli_x86_asm_macros.h:154:21: note: in expansion of macro ‘BEGIN_ASM’
  154 | #define begin_asm() BEGIN_ASM()
      |                     ^~~~~~~~~
/root/blis/kernels/zen/3/sup/bli_gemmsup_rv_zen_asm_s6x16m.c:1146:5: note: in expansion of macro ‘begin_asm’
 1146 |     begin_asm()
      |     ^~~~~~~~~
compilation terminated due to -Wfatal-errors.
[84/663] Building C object kernels/CMakeFiles/zen_KERNELS.dir/zen/3/sup/bli_gemmsup_rv_zen_asm_s6x16n.c.o
FAILED: kernels/CMakeFiles/zen_KERNELS.dir/zen/3/sup/bli_gemmsup_rv_zen_asm_s6x16n.c.o 
/usr/sbin/cc -DBLIS_IS_BUILDING_LIBRARY -DBLIS_VERSION_STRING="\"AOCL-BLAS 5.1.0 Build 20250727\"" -D_POSIX_C_SOURCE=200112L -I/root/rocm-build/blis/include/zen3 -I/root/blis/frame/3 -I/root/blis/frame/1m -I/root/blis/frame/1f -I/root/blis/frame/1 -I/root/blis/frame/include -march=native -O3 -fno-stack-protector -funroll-loops -fno-math-errno -fno-trapping-math -ffp-contract=on -ffunction-sections -fdata-sections -Wl,--gc-sections -s -w -std=c99 -fPIC -O3 -fomit-frame-pointer -mavx2 -mfpmath=sse -mfma -march=znver3 -Wall -Wno-unused-function -Wfatal-errors -fvisibility=hidden -fmacro-prefix-map=/root/blis/= -MD -MT kernels/CMakeFiles/zen_KERNELS.dir/zen/3/sup/bli_gemmsup_rv_zen_asm_s6x16n.c.o -MF kernels/CMakeFiles/zen_KERNELS.dir/zen/3/sup/bli_gemmsup_rv_zen_asm_s6x16n.c.o.d -o kernels/CMakeFiles/zen_KERNELS.dir/zen/3/sup/bli_gemmsup_rv_zen_asm_s6x16n.c.o -c /root/blis/kernels/zen/3/sup/bli_gemmsup_rv_zen_asm_s6x16n.c
In file included from /root/blis/kernels/zen/3/sup/bli_gemmsup_rv_zen_asm_s6x16n.c:39:
/root/blis/kernels/zen/3/sup/bli_gemmsup_rv_zen_asm_s6x16n.c: In function ‘bli_sgemmsup_rv_zen_asm_6x16n’:
/root/blis/frame/include/bli_x86_asm_macros.h:102:21: error: ‘asm’ operand has impossible constraints or there are not enough registers
  102 | #define BEGIN_ASM() __asm__ volatile (
      |                     ^~~~~~~
/root/blis/frame/include/bli_x86_asm_macros.h:154:21: note: in expansion of macro ‘BEGIN_ASM’
  154 | #define begin_asm() BEGIN_ASM()
      |                     ^~~~~~~~~
/root/blis/kernels/zen/3/sup/bli_gemmsup_rv_zen_asm_s6x16n.c:196:9: note: in expansion of macro ‘begin_asm’
  196 |         begin_asm()
      |         ^~~~~~~~~
compilation terminated due to -Wfatal-errors.
[94/663] Building C object kernels/CMakeFiles/zen_KERNELS.dir/zen/3/bli_trsm_small.c.o
ninja: build stopped: subcommand failed.

[AaronBeier]

You can also just build the dev branch, it builds fine now, at least for zen5 on gcc 15

[chillenb]

upstream BLIS added this in front of the affected kernels (see here):

// This avoids a known issue with GCC15 ("error: bp cannot be used in asm here", #845).
// Only check for version 15 since this may be fixed in 16 (**fingers crossed**), and also
// make sure the compiler isn't clang since it also confusingly defines __GNUC__
#if !defined(__clang__) && defined(__GNUC__) && __GNUC__ == 15
#pragma GCC optimize("-fno-tree-slp-vectorize")
#endif

Looks like the dev branch puts the same flag into the CKOPTFLAGS for the affected kernel?

[LunNova]

Fixed by a4db661